Three remarkable things that happened in AI this week: Scarlett Johansson vs. OpenAI, OpenAI vs. the media, Google vs. not running with scissors
Scarlett Johansson vs. OpenAI
What happened: Scarlett Johansson, the actress who voiced the AI in the 2013 sci-fi movie Her, accused OpenAI of using her voice without her permission. GPT-4o, released two weeks ago, can speak to users in one of five different voices, including one called “Sky” that does indeed sound like Johansson in Her. OpenAI denied that it had copied Johansson, although the fact that it had tried to hire Johansson for GPT-4o, and then CEO Sam Altman had tweeted the single word “her” shortly before 4o’s release, didn’t really help. More recently, OpenAI has explained the process by which it selected the voices for 4o, and the Washington Post reported that the actress who voiced Sky was hired months before OpenAI approached Johansson. Then OpenAI pulled the Sky voice from 4o and apologized to Johansson. So no one really knows what’s going on.
Why it matters: Most large language models (LLMs) are in a gray-to-dark-gray area when it comes to intellectual property. There simply isn’t enough high-quality training data available that is unambiguously in the public domain, leading the model builders to resort to data harvesting that either violates a platform’s terms of service or simply taking copyrighted material. This backdrop makes Johansson’s accusation far more resonant than if it had been a standalone matter (if, for example, she had accused a video game of copying her voice). Voices are indeed often protected under state-based right-to-privacy laws, although details vary by jurisdiction and it can be hard to prove intent.
OpenAI vs. traditional media
What happened: Look, I think this “vs.” framing is fun and I’m trying to stick with it, even though it doesn’t really work here. OpenAI struck a licensing deal with News Corp, the owner of the Wall Street Journal and a number of other publications. The deal not only allows OpenAI to train its models on News Corp data, but also to use its content directly–for example, by sharing a section of an article in response to a user’s prompt. The deal’s value has been cited at “possibly more than $250 million over 5 years,” according to a report in…the Journal, which is kind of weird, especially because it relies on anonymous sources. This follows similar deals with the Financial Times, Axel Springer, Reddit, and others.
Why it matters: Anyone training an LLM from scratch needs to decide whether to go full pirate for training data or not. OpenAI, which seemed the most likely to stay in the gray zone, has made a number of these deals and presumably will be looking for more. This decreases one threat (of legal action) while increasing another one–these deals are expensive, and it’s not clear whether the foundation models will bring in enough revenue to offset the eventual cost of all these deals.
Google vs. common sense
What happened: Google has started rolling out AI-created summaries at the top of many of its search results, offering a plain-language response to the user’s query in addition to its usual sponsored links. A lot of these work just fine; a lot of them don’t. Google has told users that “running with scissors is a cardio exercise that can increase your heart rate and require concentration and focus,” which, to be fair, is true. It’s also recommended changing your “blinker fluid” if your car’s blinkers aren’t making a sound, suggested adding (non-toxic) glue to pizzas to keep the cheese from sliding off, and confirmed that, yes, a dog has played in the National Hockey League.
Why it matters: You start a company, you index most of the world’s knowledge, revolutionize the way people interact with information, you more-or-less make the modern internet possible, and then you tell people to eat glue, and which part do you think they want to talk about? Like most of its big tech peers, Google has made its share of missteps with AI tools, including most recently including anachronistic diversity in the generation of historical images. Does this one matter? Probably not directly–remember in September 2012 when everyone was talking about how terrible Apple Maps was? Yeah, me neither. But Google’s AI summaries are a reminder that LLMs, like any neural-network-based tool, are inherently unpredictable. The interesting question is what balance society will strike between building more guardrails in to LLMs and simply getting used to hallucinations, the same way we don’t believe everything we read on the internet.
Three remarkable things is a more-or-less weekly roundup of interesting events from the AI community over the past more-or-less week
